Upload 6 files

Dockerfile

@@ -0,0 +1,52 @@
+FROM python:3.10-slim
+
+# 1. 环境变量
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    HF_HOME=/app/cache/huggingface \
+    PYTHONIOENCODING=UTF-8 \
+    PORT=7860 \
+    MODEL_ID="knowledgator/SMILES2IUPAC-canonical-base"
+
+# 2. 系统依赖
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl ca-certificates \
+    libstdc++6 libgomp1 \
+    libxrender1 libxext6 \
+    tini \
+ && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# 3. 用户权限
+RUN useradd -m -u 1000 user
+
+# 4. Python 依赖
+COPY requirements.txt /app/requirements.txt
+RUN pip install --upgrade pip && \
+    pip install -r requirements.txt
+
+# 5. 预下载模型
+# 必须通过 NamesConverter 加载（模型使用自定义小词表，不兼容 AutoModelForSeq2SeqLM）
+# 同时打印内部属性名，方便调试确认 hf_loader.py 中的属性探测是否正确
+RUN python -c "\
+from chemicalconverters import NamesConverter; \
+import os; \
+model_id = os.getenv('MODEL_ID', 'knowledgator/SMILES2IUPAC-canonical-base'); \
+print(f'Pre-loading: {model_id}'); \
+c = NamesConverter(model_name=model_id); \
+attrs = [a for a in dir(c) if not a.startswith('__')]; \
+print(f'NamesConverter attrs: {attrs}'); \
+print('Pre-load complete.')"
+
+# 6. 复制文件
+COPY --chown=user:user . /app
+RUN mkdir -p /app/cache/huggingface && chown -R user:user /app/cache
+
+# 7. 启动
+USER user
+EXPOSE $PORT
+ENTRYPOINT ["/usr/bin/tini", "--"]
+CMD ["sh", "-c", "uvicorn app:app --host 0.0.0.0 --port ${PORT:-7860}"]

README.md

@@ -1,10 +1,96 @@
 ---
-title: MolScribe
-emoji: 👁
-colorFrom: purple
-colorTo: red
+title: SMILES → IUPAC (Docker)
+emoji: 🧪
+colorFrom: indigo
+colorTo: blue
 sdk: docker
 pinned: false
 ---
+# SMILES→IUPAC(Attention-Optimized V6.6)
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+基于`knowledgator/SMILES2IUPAC-canonical-base`模型的SMILES转IUPAC高精度转换服务。本项目集成了FastAPI与Gradio，引入了基于注意力的跨度预优化、动态束搜索与拓扑幻觉拦截机制，完美兼顾了简单芳香环的高效解码与复杂手性骨架的深度探索。
+
+## 核心特性
+
+* **动态束搜索(Dynamic Beam Search)**：系统会通过RDKit实时计算分子的重原子数量(Heavy Atoms)。小分子(如苯环，≤10)自动采用贪婪解码(`beams=1`)彻底根治同位素幻觉；中型分子(≤20)使用`beams=4`平衡速度；大分子(如福莫特罗，>20)自动开启深度搜索(`beams=8`)防止长序列推理断连。
+* **TTA短路验证机制(Short-circuit Evaluation)**：在处理复杂大分子时，系统会生成包含50个合规变体的候选池，并依次送入大语言模型进行推理。一旦某个变体生成的IUPAC名称通过了拓扑安全校验，即刻短路跳出并返回结果，极大节省了显存算力与API响应时间。
+* **纯随机遍历预优化**：彻底废除了破坏底层芳香属性的强制Kekulize操作，恢复极具多样性的纯随机遍历(`doRandom=True`)算法，并配合生成大写凯库勒式(`kekuleSmiles=True`)，在最大化变体结构多样性的同时，完美契合大模型底层的无芳香标志预训练分布。
+* **RDKit拓扑拦截过滤**：基于真实分子拓扑结构，自动拦截并剔除AI生成的含有不存在环系(如`cyclohept`、`cyclooct`、`cyclonon`)的虚假名称。
+* **双重交互模式**：提供标准化RESTful API供程序调用，同时挂载Gradio WebUI方便直观调试与查看变体推理细节。
+
+## 环境变量配置
+
+* `DISABLE_CANONICALIZE`：设为`1`或`true`可全局关闭服务端强制规范化(不推荐，关闭后若输入非标SMILES可能诱发幻觉)。
+* `TTA_SAMPLES`：TTA候选池的评估上限，开启`use_tta`时生效，默认最少评估`5`个优质变体。
+* `MODEL_ID`：底层模型路径，默认为`knowledgator/SMILES2IUPAC-canonical-base`。
+
+## REST API调用指南
+
+### 1.服务健康检查
+**GET** `/healthz`
+返回结果：
+```json
+{
+  "ok": true
+}
+
+```
+
+### 2.单条SMILES转换
+
+**POST** `/api/smiles2iupac`
+**请求体(JSON)：**
+
+```json
+{
+  "smiles": "COc1ccc(CC(C)NCC(O)c2ccc(O)c(NC=O)c2)cc1",
+  "canonicalize": true,
+  "style": "BASE",
+  "use_tta": true
+}
+
+```
+
+**字段说明：**
+
+* `smiles`：需要转换的SMILES字符串。
+* `canonicalize`：是否允许服务端进行标准化处理(强烈建议为`true`)。
+* `style`：命名风格，支持`BASE`(默认推荐，兼顾俗名与系统名)、`SYST`(纯系统命名)、`TRAD`(传统命名)。
+* `use_tta`：是否开启跨度预优化与短路验证机制。
+
+**响应体(JSON)：**
+
+```json
+{
+  "success": true,
+  "input": "COc1ccc(CC(C)NCC(O)c2ccc(O)c(NC=O)c2)cc1",
+  "style": "BASE",
+  "tta_used": true,
+  "candidates_count": 5,
+  "iupac": "N-[2-hydroxy-5-[1-hydroxy-2-[[1-(4-methoxyphenyl)propan-2-yl]amino]ethyl]phenyl]formamide",
+  "voting_details": {
+    "N-[2-hydroxy-5-[1-hydroxy-2-[[1-(4-methoxyphenyl)propan-2-yl]amino]ethyl]phenyl]formamide": 1
+  }
+}
+
+```
+
+### 3.批量SMILES转换
+
+**POST** `/api/smiles2iupac/batch`
+**请求体(JSON)：**
+
+```json
+{
+  "inputs": [
+    {
+      "smiles": "c1ccccc1",
+      "style": "BASE"
+    },
+    {
+      "smiles": "CC(=O)Oc1ccccc1C(=O)O",
+      "style": "SYST",
+      "use_tta": true
+    }
+  ]
+}

app.py

@@ -0,0 +1,268 @@
+# -*- coding: utf-8 -*-
+import os
+import re
+from collections import Counter
+import gradio as gr
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+from typing import List, Optional, Union
+
+from rdkit import Chem
+from rdkit.Chem import MolToSmiles # 彻底移除了破坏性的Kekulize
+
+from hf_loader import smiles2iupac
+
+app = FastAPI(
+    title="SMILES→IUPAC(Attention-Optimized)",
+    version="6.6.0",
+)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+def calculate_max_ring_span(smiles: str) -> int:
+    def repl(m):
+        return '_' * len(m.group())
+    
+    clean_smiles = re.sub(r'\[.*?\]', repl, smiles)
+    
+    active_rings = {}
+    max_span = 0
+    
+    for match in re.finditer(r'%\d{2}|\d', clean_smiles):
+        ring_id = match.group()
+        pos = match.start()
+        
+        if ring_id in active_rings:
+            span = pos - active_rings[ring_id]
+            if span > max_span:
+                max_span = span
+            del active_rings[ring_id]
+        else:
+            active_rings[ring_id] = pos
+            
+    return max_span
+
+def generate_optimized_smiles(s: str, pool_size: int = 50, top_k: int = 5) -> List[str]:
+    mol = Chem.MolFromSmiles(s)
+    if mol is None:
+        raise ValueError(f"非法SMILES，RDKit无法解析:{s!r}")
+
+    # 获取标准规范凯库勒式作为保底(RDKit内部会自动处理，不破坏mol对象)
+    canonical_kekule = MolToSmiles(mol, canonical=True, kekuleSmiles=True)
+    num_atoms = mol.GetNumAtoms()
+
+    if num_atoms < 15:
+        return [canonical_kekule]
+
+    tta_set = {canonical_kekule}
+    
+    # 核心修复：恢复极其成功的随机算法doRandom=True
+    max_attempts = pool_size * 5
+    attempts = 0
+    while len(tta_set) < pool_size and attempts < max_attempts:
+        try:
+            variant = MolToSmiles(mol, canonical=False, doRandom=True, kekuleSmiles=True)
+            tta_set.add(variant)
+        except Exception:
+            pass
+        attempts += 1
+
+    scored_smiles = []
+    for sm in tta_set:
+        span_score = calculate_max_ring_span(sm)
+        scored_smiles.append((span_score, len(sm), sm))
+
+    scored_smiles.sort()
+    return [item[2] for item in scored_smiles[:top_k]]
+
+
+DISABLE_CANONICALIZE = os.getenv("DISABLE_CANONICALIZE", "").lower() in ("1", "true", "yes")
+TTA_SAMPLES = int(os.getenv("TTA_SAMPLES", "5"))
+
+class SMILESItem(BaseModel):
+    smiles: str
+    canonicalize: Optional[bool] = True
+    style: Optional[str] = "BASE"
+    use_tta: Optional[bool] = True
+    beams: Optional[Union[int, str]] = "auto"
+
+class BatchRequest(BaseModel):
+    inputs: List[SMILESItem]
+
+def process_single_smiles(s: str, do_canon: bool, style: str, use_tta: bool, beams: Union[int, str] = "auto"):
+    mol = Chem.MolFromSmiles(s)
+    if mol is None:
+        return "", [s], {}, 1
+
+    if str(beams).lower() == "auto":
+        heavy_atoms = mol.GetNumHeavyAtoms()
+        if heavy_atoms <= 10:
+            dynamic_beams = 1  
+        elif heavy_atoms <= 20:
+            dynamic_beams = 4  
+        else:
+            dynamic_beams = 8  
+    else:
+        try:
+            dynamic_beams = int(beams)
+            if dynamic_beams < 1: 
+                dynamic_beams = 1
+        except ValueError:
+            dynamic_beams = 4
+
+    ring_info = mol.GetRingInfo().AtomRings()
+    ring_sizes = set(len(r) for r in ring_info)
+    
+    hallucination_blacklist = []
+    if 7 not in ring_sizes: hallucination_blacklist.append("cyclohept")
+    if 8 not in ring_sizes: hallucination_blacklist.append("cyclooct")
+    if 9 not in ring_sizes: hallucination_blacklist.append("cyclonon")
+
+    if not do_canon:
+        try:
+            # 同样移除对mol对象的破坏，直接输出大写
+            s_kekule = Chem.MolToSmiles(mol, canonical=False, kekuleSmiles=True)
+        except Exception:
+            s_kekule = s
+            
+        name = smiles2iupac(s_kekule, style=style, num_beams=dynamic_beams)
+        return name, [s_kekule], {name: 1}, dynamic_beams
+
+    sample_count = TTA_SAMPLES if use_tta else 1
+    if use_tta and sample_count < 5:
+        sample_count = 5 
+        
+    smiles_list = generate_optimized_smiles(s, pool_size=50, top_k=sample_count)
+    
+    evaluated_smiles = []
+    best_name = ""
+    
+    for current_smiles in smiles_list:
+        evaluated_smiles.append(current_smiles)
+        name = smiles2iupac(current_smiles, style=style, num_beams=dynamic_beams)
+        
+        if not name:
+            continue
+            
+        name_lower = name.lower()
+        if any(bad_word in name_lower for bad_word in hallucination_blacklist):
+            continue
+            
+        best_name = name
+        break
+
+    if not best_name:
+        return "未能生成符合拓结构的名称(全被过滤器拦截)", evaluated_smiles, {}, dynamic_beams
+
+    return best_name, evaluated_smiles, {best_name: 1}, dynamic_beams
+
+@app.get("/healthz")
+def healthz():
+    return {"ok": True}
+
+@app.post("/api/smiles2iupac")
+def api_smiles2iupac(req: SMILESItem):
+    try:
+        s = (req.smiles or "").strip()
+        if not s:
+            return {"success": False, "error": "输入为空"}
+
+        do_canon = (req.canonicalize if req.canonicalize is not None else True) and (not DISABLE_CANONICALIZE)
+        valid_styles = ["BASE", "SYST", "TRAD"]
+        style = req.style.upper() if req.style and req.style.upper() in valid_styles else "BASE"
+        use_tta = req.use_tta if req.use_tta is not None else True
+        beams = req.beams if req.beams is not None else "auto"
+
+        best_name, smiles_list, counts, _ = process_single_smiles(s, do_canon, style, use_tta, beams)
+
+        return {
+            "success": True,
+            "input": s,
+            "style": style,
+            "tta_used": use_tta,
+            "candidates_count": len(smiles_list),
+            "iupac": best_name,
+            "voting_details": counts
+        }
+    except Exception as e:
+        return {"success": False, "error": str(e)}
+
+@app.post("/api/smiles2iupac/batch")
+def api_smiles2iupac_batch(req: BatchRequest):
+    out = []
+    for item in req.inputs:
+        try:
+            s = (item.smiles or "").strip()
+            if not s:
+                out.append({"success": False, "error": "输入为空"})
+                continue
+
+            do_canon = (item.canonicalize if item.canonicalize is not None else True) and (not DISABLE_CANONICALIZE)
+            valid_styles = ["BASE", "SYST", "TRAD"]
+            style = item.style.upper() if item.style and item.style.upper() in valid_styles else "BASE"
+            use_tta = item.use_tta if item.use_tta is not None else True
+            beams = item.beams if item.beams is not None else "auto"
+
+            best_name, smiles_list, counts, _ = process_single_smiles(s, do_canon, style, use_tta, beams)
+            
+            out.append({
+                "success": True,
+                "input": s,
+                "style": style,
+                "iupac": best_name
+            })
+        except Exception as e:
+            out.append({"success": False, "input": item.smiles, "error": str(e)})
+    return out
+
+def gradio_fn(s: str, style: str, canonicalize: bool, use_tta: bool, beams: str):
+    if not (s or "").strip():
+        return "", "输入为空"
+    try:
+        do_canon = canonicalize and (not DISABLE_CANONICALIZE)
+        best_name, smiles_list, counts, dynamic_beams = process_single_smiles(s, do_canon, style, use_tta, beams)
+        
+        if not do_canon and not use_tta:
+            debug_info = (
+                f"原始输入:{s}\n"
+                f"[直通模式]:动态BeamSearch宽度:{dynamic_beams}\n\n"
+                f"模型直接输出:{best_name}"
+            )
+        else:
+            variants_text = "\n".join([f"  - {x}" for x in smiles_list])
+            attempts_count = len(smiles_list)
+            
+            debug_info = (
+                f"原始输入:{s}\n"
+                f"动态BeamSearch宽度:{dynamic_beams}\n\n"
+                f"执行短路验证(大模型实际推理了{attempts_count}个高质量变体):\n{variants_text}\n\n"
+                f"最终采纳结果(验证通过):\n  - {best_name}\n"
+            )
+        return best_name, debug_info
+    except Exception as e:
+        return "", f"Error:{e}"
+
+demo = gr.Interface(
+    fn=gradio_fn,
+    inputs=[
+        gr.Textbox(label="输入SMILES", placeholder="支持任意合法SMILES写法"),
+        gr.Radio(["BASE", "SYST", "TRAD"], label="命名风格", value="BASE"),
+        gr.Checkbox(label="自动Kekulé规范化(强烈建议开启)", value=True),
+        gr.Checkbox(label="开启跨度预优化与拓扑过滤(解决复杂环系幻觉)", value=True),
+        gr.Dropdown(["auto", "1", "4", "8", "16"], label="BeamSearch宽度(选auto为智能调整)", value="auto", allow_custom_value=True),
+    ],
+    outputs=[
+        gr.Textbox(label="最优IUPAC名称", interactive=True),
+        gr.Textbox(label="调试信息", lines=12),
+    ],
+    title="SMILES→IUPAC(Attention-Optimized V6.6)",
+    description="恢复基于纯随机遍历的变体生成算法以最大化结构多样性；配合动态束搜索与短路拦截，实现精度与速度的双赢。",
+)
+
+app = gr.mount_gradio_app(app, demo, path="/")

gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

hf_loader.py

@@ -0,0 +1,40 @@
+# -*- coding: utf-8 -*-
+import os
+from functools import lru_cache
+from typing import Union, List
+from chemicalconverters import NamesConverter
+
+MODEL_ID = os.getenv("MODEL_ID", "knowledgator/SMILES2IUPAC-canonical-base")
+
+@lru_cache(maxsize=1)
+def _load() -> NamesConverter:
+    print(f"[hf_loader]Loading:{MODEL_ID}...")
+    converter = NamesConverter(
+        model_name=MODEL_ID,
+        smiles_max_len=512,
+        iupac_max_len=512
+    )
+    print("[hf_loader]Loaded.")
+    return converter
+
+def smiles2iupac(smiles_kekule: Union[str, List[str]], style: str = "BASE", num_beams: int = 4) -> Union[str, List[str]]:
+    """
+    支持单条或批量SMILES转换为IUPAC名称，支持动态束搜索宽度
+    """
+    valid_styles = {"BASE", "SYST", "TRAD"}
+    style_upper = style.upper() if style and style.upper() in valid_styles else "BASE"
+    
+    converter = _load()
+    is_list = isinstance(smiles_kekule, list)
+    smiles_list = smiles_kekule if is_list else [smiles_kekule]
+
+    results = []
+    for s in smiles_list:
+        input_text = f"<{style_upper}>{s}"
+        res = converter.smiles_to_iupac(input_text, num_beams=num_beams)
+        if isinstance(res, list):
+            results.append(res[0] if res else "")
+        else:
+            results.append(str(res))
+
+    return results if is_list else results[0]

requirements.txt

@@ -0,0 +1,13 @@
+fastapi
+uvicorn
+gradio>=4.0
+huggingface_hub
+# 官方库 — 必须保留，因为模型使用自定义小词表（encoder:137 / decoder:822 tokens）
+# 只有通过 NamesConverter 才能正确加载这两个自定义 tokenizer
+# 我们在 hf_loader.py 中绕过了其内部有 bug 的预处理，直接操控底层推理
+chemical-converters>=0.1.2
+transformers>=4.35.0
+torch --extra-index-url https://download.pytorch.org/whl/cpu
+sentencepiece
+protobuf
+rdkit-pypi